svn commit: r210586 - projects/ofed/head/sys/ofed/include/linux

[Alan Cox]

Jeff Roberson wrote:
> On Wed, 28 Jul 2010, Alan Cox wrote:
>
>> Jeff Roberson wrote:
>>> Author: jeff
>>> Date: Thu Jul 29 02:12:22 2010
>>> New Revision: 210586
>>> URL: http://svn.freebsd.org/changeset/base/210586
>>>
>>> Log:
>>>    - Refine various page allocation methods.  They all essentially 
>>> override
>>>      the page object pointer to cache the virtual address after 
>>> allocation.
>>>      This same trick is used in UMA to store the slab address and so it
>>>      should be safe.
>>>
>>
>> You needn't do this.  The virtual address of a kernel_object or 
>> kmem_object page is easily obtained:
>>
>>   VM_MIN_KERNEL_ADDRESS + IDX_TO_OFF(m->pindex)
>
> Ah, of course.  I hadn't considered that.  Thanks.  I'll remove the 
> other hackery.
>
> On another note; linux guarantees multipage allocations are aligned on 
> the first power of two size that satisfies the allocation.  I can do 
> this with contigmalloc but they also guarantee that the virtual 
> address is similarly aligned.  I don't think I've run across anything 
> that requires it, but if I do, do we have any provision for allocating 
> virtually aligned addresses?
>

No, I'm afraid not.  We only support a couple of specific cases of 
virtually aligned allocation: thread stacks on MIPS and the start of the 
large submaps on superpage-enabled architectures.

Alan